Building Multipurpose Systems With Dynamic Function Exchange Part Three: Tools For Deploying DFX

In our previous two articles, we introduced you to the concept of Dynamic Function Exchange (DFX), a design approach that dynamically reallocates unused system resources to other tasks. We also introduced you to some techniques for bundling and managing your DFX resources. In this article, we will discuss some of the adaptive computing tools that make DFX possible.

As part of its investment in bringing adaptive computing to a wide range of new applications, Xilinx has packaged its adaptive computing technology into a PCI form factor. Alveo is a flexible module platform that reconfigures hardware to become the optimal processing resources needed for each application (see figure 1).

Fig. 1: Alveo is a flexible module platform that reconfigures hardware to become the optimal processing resources needed for each application.

Alveo is much more than just a combination of compute, storage, and network resources. It has a full solution stack and is supported by the Vitis unified software development platform. Vitis enables developers to leverage the efficiency of programmable logic without complexity and to work at a high level, much like developing using C++ (see figure 2). Xilinx has also developed libraries providing production-ready components to help developers get to market faster. Alveo PCI cards are production-ready and must pass a rigorous qualification process so enterprises can quickly deploy them with confidence.

Fig. 2: A unified software development platform like Vitis (shown) enables developers to leverage the efficiency of programmable logic without complexity and to work at a high level, much like developing using C++.

With the flexibility of DFX and Alveo, servers can be dynamically reconfigured to meet real-time capacity requirements. Data centers are expensive to build and manage. They also have space and power constraints. Ideally, IT needs to provide the highest output per square foot and power budget. DFX enables the same server to be heavily receive-based or transmit-based, depending upon the application and current need. For example, if there is a large increase in ingest traffic, the Alveo can be reconfigured to accommodate this. When traffic decreases, the same Alveo card used to ingest traffic can be configured to distribute video to thousands of endpoints.

A single, small Alveo adaptive computing card can replace a full dual socket high-end server while consuming 1000% lower power and delivering video at a 5X lower cost per channel. The high capacity of the Alveo also increases video capacity per server and rack.

There are other savings as well. Typically, each SoC or ASIC must have its own power, clock, and memory. And while these devices can be placed standby when they are not needed, they still consume power and space. Thus, cost savings can be calculated at a module level (footprint, components, and power). Operational savings can be accounted for as well. Finally, fewer components reduces the number of points of failure, leading to increased reliability.

DFX and Alveo can also be used to enable data centers to adapt to the rapid evolution of applications and data brought about by AI. As AI services proliferate, this has increased the amount of unstructured data available for use to gain better insight in operations and how to make better decisions. Traditional storage is not enough for these applications. Computational storage, or smart storage, is critical to enabling the efficient processing of these new workloads.

With Alveo, intelligence can be brought closer to storage. For example, as video is sent to storage, it can be encoded without loading a server CPU. In addition, when the video is retrieved, Alveo can transcode the stream as required by the application, again without loading a server CPU.

For applications where hardware acceleration is essential, an adaptive computing platform combined with DFX provides a flexible architecture that optimizes for performance, efficiency, and utilization.